The invention addresses the problem of providing an image labeling system, which allows reducing the human intervention in labeling procedure.
| Patent title | Image labeling system |
|---|---|
| Thematic area | Industry, Digital and Security |
| Ownership | ALMA MATER STUDIORUM - UNIVERSITA' DI BOLOGNA |
| Inventors | Daniele De Gregorio, Luigi Di Stefano |
| Protection | Italy, Europe, USA, China, Hong Kong |
| Licensing status | Licensed |
| Keywords | Convolutional neural networks, Deep networks, Dataset generation, Labeled data |
| Filed on | 18 April 2019 |
Machine Learning systems, such as Neural Networks, have to be trained with data and the availability of such data is crucial. In this context, images are fundamental for several data-driven models and these images to be used to train such models are hand-labeled by humans. Converting to an automated configuration seems to be the optimal solution.
The system is conceived for the automated generation of Labeled Data for training Data-Driven Vision systems, such as the Neural Networks. The patented technology allows to generate automatically a quantity of visual data (e.g. labeled images) enough to be able to train, and/or reconfigure a vision system, for the recognition of a set of objects of interest. In contrast with the conventional systems, once the operational setup is changed (when objects of interest, or system positioning, or, or when lights ta are changed, etc), there is no need for manual reconfiguration. The device thus consists of an automatic machine with multiple degrees of freedom (e.g. an anthropomorphic robot) and a display, such as a simple LCD or E-INK monitor, capable of presenting information in visual form.
APPLICATION:
- Industry and manufacturing;
- Augmented Reality.
ADVANTAGES:
- Lowering costs;
- Reducing human involvement throughout the labelling process.